R76.H5 — Audit: colibri-sigma-orchestrator skill stub

Intent

R76.H5 creates the missing .agents/skills/colibri-sigma-orchestrator/SKILL.md file (plus its .claude/skills/ mirror) so the T1 Sigma orchestrator role has an executable skill entry matching its contract at docs/agents/sigma-orchestrator.md. The skill is a contract-reference pointer, not a tool-dispatch runtime (agent runtime is deferred to Phase 1.5 per ADR-005 §Decision).

Surface inventory

1. Sigma contract (authoritative, read-only)

• Path: docs/agents/sigma-orchestrator.md
• Length: 262 lines (verified this audit)
• Structure (9 sections):
  1. Scope of Authority (may / may not)
  2. Phase A — Plan (manifest template + writeback)
  3. Phase W — Waves (wave gate protocol)
  4. Phase B — Seal (audit_verify_chain + merkle_finalize + seal doc)
  5. Interaction with Other Tiers (T0 / T2 / T3)
  6. Artifacts Sigma Produces (table)
  7. Per-Round Start Checklist
  8. Per-Round Seal Checklist
  9. Non-Negotiables (6 rules)
• Authority: authoritative. Not to be rewritten under this task — only referenced.

2. Target skill path — canon

• Path: .agents/skills/colibri-sigma-orchestrator/SKILL.md
• Status: does not exist (verified — ls returned No such file or directory)
• Directory .agents/skills/colibri-sigma-orchestrator/ does not exist either — must be created.

3. Target skill path — mirror

• Path: .claude/skills/colibri-sigma-orchestrator/SKILL.md
• Status: does not exist (verified — ls returned No such file or directory)
• Directory .claude/skills/colibri-sigma-orchestrator/ does not exist either — must be created.
• Per CLAUDE.md §9.2, .claude/skills/colibri-* is a known mirror of .agents/skills/colibri-*; editing by hand is forbidden except on first creation (this task), where canon + mirror must be byte-identical.

4. Pattern skill — colibri-pm/SKILL.md

• Path: .agents/skills/colibri-pm/SKILL.md
• Length: 263 lines (verified)
• Role: the T2 contract-reference skill — the closest analogue to what R76.H5 must create.
• Structural template (adapted for Sigma below):
  • YAML frontmatter: name, description (quoted, multiple colons), round, updated
  • H1 title + greek-letter + tier + phase-0 reality stamp blockquote
  • Session startup protocol (numbered list of tool invocations)
  • FSM / phase-model section
  • Dispatch pattern section
  • Writeback protocol section
  • Common patterns / typical requests
  • References (CLAUDE.md sections + contract)

5. Live Sigma artifact — R76 spawn manifest

• Path: .agents/spawns/r76-phase-0-to-1-bridge/manifest.md
• Status: exists on main as of PR #152
• Purpose: demonstrates Sigma’s Phase A output for R76 itself — the first round to actually populate .agents/spawns/rNN-<theme>/ per the contract.
• This is the concrete example the skill should point readers to when asking “what does a Phase A manifest look like in practice”.

Gap

The Sigma contract is a 262-line document. Without an executable skill file:

• Claude sessions cannot discover the Sigma role via the skill registry (skill_list).
• T0 users cannot trigger Phase A / B workflow via “open round R76” or “seal R76” prompts.
• PM / Executor skills reference the contract but no peer skill exists to link back to them at Sigma’s tier.

Scope

In-scope

1. Create .agents/skills/colibri-sigma-orchestrator/SKILL.md as a contract-reference skill stub.
2. Mirror canon into .claude/skills/colibri-sigma-orchestrator/SKILL.md byte-identical on creation.
3. Add 5-step chain artifacts (audit, contract, packet, verification).
4. Pass npm run build && npm run lint && npm test gates.

Out-of-scope

• Rewriting docs/agents/sigma-orchestrator.md (authoritative; only reference).
• Creating Sigma MCP tools (no agent runtime in Phase 0; Phase 1.5+ per ADR-005).
• Editing src/ (skill stub touches .agents/ + .claude/ + docs/ only).
• Editing other skills (R76.H4 handles inventory-r74-2 ref cleanup; no collision).

Risks

1. YAML hazard (memory-tracked): the description: field contains multiple colons (Phase A:, Greek:, Contract:). Must be double-quoted to survive gray-matter parsing. Verified via post-write head -20.
2. Canon/mirror divergence: byte-identical on creation is mandatory. Verify via diff in Step 5.
3. Overclaim risk: the skill must NOT imply Sigma has executable dispatch authority via an MCP tool. Phase 0 reality stamp must explicitly note agent runtime is Phase 1.5+.
4. Line budget: task prompt requires under 200 lines. Pattern skill colibri-pm is 263 lines; Sigma skill must be compressed via references rather than duplication.

References

• CLAUDE.md §2 (tier hierarchy), §3 (worktree), §7 (writeback)
• docs/agents/sigma-orchestrator.md (authoritative contract — 262 lines)
• .agents/skills/colibri-pm/SKILL.md (structural template — 263 lines)
• .agents/spawns/r76-phase-0-to-1-bridge/manifest.md (live Phase A artifact)
• ADR-005 §Decision (agent runtime deferral)
• Memory: feedback_yaml_frontmatter_hazard.md (quote description strings with colons)